AI-103 Practice Questions: Implement generative AI and agentic solutions Domain

Correct answer (B): RAG is the best fit when an app must answer from frequently changing enterprise content. By retrieving current HR documents at runtime and grounding the response on those results, the assistant can stay up to date without retraining the model whenever policies change.

Why the other options are wrong:
- Option A: Fine-tuning is not the primary solution for keeping an assistant current with changing business documents. Retrieval over an up-to-date knowledge source is usually the better fit.
- Option C: Content safety filters can reduce harmful output, but they do not provide grounding or keep answers aligned to current HR policy content.
- Option D: A larger model does not reliably solve hallucination or freshness problems. Improving retrieval and grounding is usually more effective.

Correct answer (A): The symptom points to retrieval problems, not generation settings. Missing newly uploaded contracts and weak citations indicate issues such as chunking, embedding strategy, metadata filtering, or index freshness. Those retrieval factors are the best place to investigate first.

Why the other options are wrong:
- Option B: Content filters address harmful or policy-violating output, not whether the right documents are retrieved and cited.
- Option C: Fine-tuning on older samples does not solve stale or missing source retrieval. The problem described is missing current evidence, not lack of model adaptation.
- Option D: Prompt reminders may help formatting, but they do not fix retrieval failures caused by chunking, embeddings, filters, or index freshness.

Correct answer (B): Safety filtering and grounding solve different problems. Content safety filters help reduce harmful or disallowed output, but they do not guarantee factual grounding. To reduce unsupported answers in a knowledge assistant, the better next step is to improve retrieval quality, grounding strategy, and source citation behavior.

Why the other options are wrong:
- Option A: Safety filter strictness does not guarantee factual grounding or source-backed answers.
- Option C: Fine-tuning on transcripts is not the primary fix for unsupported answers when the core issue is weak grounding to enterprise sources.
- Option D: Disabling retrieval removes grounding and usually makes unsupported answers more likely, not less.

Correct answer (B): This is the best design because the scenario combines two requirements: human review for a high-risk action and least-privilege access for the agent's tool. The blueprint explicitly states that high-risk actions should use human-in-the-loop review or explicit confirmation, and that least-privilege access applies to agent tools as well as Azure resources.

Why the other options are wrong:
- Option A: Content filtering helps reduce unsafe outputs, but it does not replace authorization controls or a required human approval step for sensitive transactions.
- Option C: Full purchasing API access violates least-privilege design and increases the blast radius of mistakes or misuse.
- Option D: A larger model does not remove the need for workflow controls, human approval, or scoped permissions for sensitive actions.

Correct answer (C): Live shipment status is transactional data held in a system of record. A tool is a better choice than RAG when the app must retrieve live data or perform an action against such a system. Structured tool or function calling lets the agent query the ERP at runtime in a controlled way.

Why the other options are wrong:
- Option A: Weekly exports create stale data and are not the best fit for live transactional status checks.
- Option B: Fine-tuning does not keep the model current with live order state and is not an appropriate substitute for querying the source system.
- Option D: Semantic ranking on policy content cannot retrieve the current state of a specific order from the ERP system.

Correct answer (B): Tool or function calling is the right pattern when a model must invoke APIs with structured arguments. It is more reliable than prompt-only text when the app needs deterministic tool selection and parameter formatting.

Why the other options are wrong:
- Option A: More examples can help somewhat, but prompt text alone is less reliable for structured API invocation than tool calling with a schema.
- Option C: Fine-tuning is not the primary pattern for invoking business APIs with well-formed parameters.
- Option D: Lowering temperature may reduce variation, but it does not provide the structured contract enforcement that tool calling is designed for.

Correct answer (B): Refunds are high-impact and potentially irreversible. In that scenario, the best practice is to include a human approval step instead of allowing fully autonomous execution. This adds governance and reduces risk for sensitive actions.

Why the other options are wrong:
- Option A: Model confidence alone is not an adequate control for risky or irreversible business actions.
- Option C: Fine-tuning on historical behavior does not remove the governance need for approval on sensitive actions.
- Option D: Content safety filters address harmful content, not authorization or approval for financial transactions.

Correct answer (C): This scenario needs both semantic similarity for paraphrased questions and text matching for exact codes or SKUs. Azure AI Search hybrid search is designed for that combination, and vector retrieval requires embedding the content and queries into vectors stored in vector fields.

Why the other options are wrong:
- Option A: Semantic ranking can improve ranking over textual content, but it does not replace vector retrieval when semantic similarity across phrasing is required.
- Option B: Keyword search can help with exact terms, but it is weaker for paraphrased or semantically similar questions. The scenario requires both behaviors.
- Option D: Fine-tuning the model does not replace a search design that can match both exact terms and semantic similarity.

Correct answer (B): Deploying a model and deploying the application are separate concerns. The app must be versioned and configured to use the intended model deployment, prompts, tools, and related assets. Production readiness also includes evaluation before rollout and monitoring after deployment. For agentic workflows, tracing is especially useful to diagnose which prompt, retrieval step, tool call, or policy decision caused slow or incorrect behavior.

Why the other options are wrong:
- Option A: A model deployment does not automatically update application behavior. The app still needs its own release and configuration changes.
- Option C: Content safety filtering does not manage rollout state, app configuration, or multi-step execution diagnostics.
- Option D: A larger model does not address the core issues here, which are release management and observability.

Correct answer (C): Hybrid search is the best answer because the scenario requires both exact lexical matching for identifiers like part numbers and embedding-based similarity for conceptually similar phrasing. The blueprint specifically identifies hybrid retrieval as the common choice when both lexical and vector-style matching are needed.

Why the other options are wrong:
- Option A: Vector search is useful for semantic similarity, but by itself it is not the best fit when exact lexical matches are also critical.
- Option B: Keyword search can help with exact identifiers, but it may miss semantically similar content when wording differs from the source documents.
- Option D: Semantic search can improve meaning-aware retrieval or ranking, but the scenario explicitly calls for both lexical and embedding-based matching, which is why hybrid is the best choice here.